Fetal brain age prediction network training method, application method and electronic equipment

Abstract

The invention provides a fetal brain age prediction network training method, an application method and electronic equipment, and belongs to the technical field of medical image processing, and the training method comprises the steps: carrying out the similarity sorting conversion of a fetal brain magnetic resonance image and a corresponding brain age label, and obtaining a similarity sorting matrix, performing vector embedding reconstruction on the similarity sorting matrix to obtain an embedded vector, and determining an initial permutation entropy; performing median absolute deviation weighting on the initial permutation entropy to obtain a weighted permutation entropy, performing multi-scale ordinal number information reconstruction on the weighted permutation entropy to obtain a multi-scale permutation entropy, performing time sequence enhancement reconstruction on the multi-scale permutation entropy to obtain a time sequence multi-scale permutation entropy, and constructing permutation entropy regularization loss according to the time sequence multi-scale permutation entropy; and obtaining a fetal brain age prediction network according to permutation entropy regularization loss iterative training. According to the method, the loss function is constructed through the permutation entropy regularizer, so that the continuous ordered relation in regression can be captured, and the distinguishing ability and generalization ability of the fetal brain age prediction network are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a fetal brain age prediction network training method, an application method and an electronic device. Background Art

[0002] Predicting fetal brain age based on sulcal and gyral development has been widely used to characterize normal fetal brain development and variation. Changes in fetal brain age are closely linked to the pathogenesis of neurological diseases. The discrepancy between predicted and actual brain age can serve as an indicator of deviation from the typical developmental trajectory, providing important data support for prenatal screening and early diagnosis of congenital anomalies.

[0003] However, in existing fetal brain age prediction methods, the focus during the model training phase is usually on directly learning the mapping relationship required from input to output, without considering the important continuous and ordered representation in regression (for example, magnetic resonance imaging of fetal brains of similar brain ages often show high similarity). As a result, the learned representations often exhibit harmful discretization, destroying the inherent continuity and ordinal relationship between regression samples, weakening the model's discrimination and generalization capabilities, and affecting the accuracy of fetal brain age prediction.

[0004] Therefore, the existing technology has the technical problem of not considering the continuous ordered relationship in regression, resulting in a decrease in the model's discrimination and generalization capabilities, and low accuracy in predicting fetal brain age, which needs to be improved. Summary of the Invention

[0005] In view of this, it is necessary to provide a fetal brain age prediction network training method, application method and electronic equipment to capture the continuous ordered representation in regression, enhance the model's discrimination and generalization capabilities, and improve the accuracy of brain age prediction.

[0006] In a first aspect, the present invention provides a fetal brain age prediction network training method, comprising: Perform similarity ranking transformation on the fetal brain magnetic resonance images and the corresponding brain age labels to obtain a similarity ranking matrix, perform vector embedding reconstruction on the similarity ranking matrix to obtain an embedding vector, and determine the initial permutation entropy of the embedding vector; The initial permutation entropy is weighted by the median absolute deviation to obtain the weighted permutation entropy, the weighted permutation entropy is reconstructed by multi-scale ordinal information to obtain the multi-scale permutation entropy, the multi-scale permutation entropy is reconstructed by time series enhancement to obtain the time series multi-scale permutation entropy, and the permutation entropy regularization loss is constructed based on the time series multi-scale permutation entropy; The initial fetal brain age prediction network was iteratively trained according to the permutation entropy regularization loss to obtain the fetal brain age prediction network.

[0007] In some possible implementations, the fetal brain magnetic resonance image is obtained by preprocessing an initial magnetic resonance image, and the preprocessing of the initial magnetic resonance image includes: The initial magnetic resonance image is subjected to key region cutting, histogram equalization, image transformation and zero filling operations in sequence.

[0008] In some possible implementations, the similarity ranking matrix includes a label ranking matrix and a feature calibration matrix. The similarity ranking matrix is obtained by performing similarity ranking transformation on the fetal brain magnetic resonance images and the corresponding brain age labels, including: A similarity matrix is constructed for the fetal brain magnetic resonance images and the corresponding brain age labels to obtain a feature similarity matrix and a label similarity matrix; Arrange the label similarity matrix in ascending order to obtain a label sorting matrix, and determine the sorting conversion rule based on the label sorting matrix; The feature similarity matrix is calibrated according to the sorting transformation rule to obtain the feature calibration matrix.

[0009] In some possible implementations, performing vector embedding reconstruction on the similarity ranking matrix to obtain an embedded vector and determining an initial permutation entropy of the embedded vector include: Perform delayed embedding state reconstruction on the similarity ranking matrix to obtain the embedding vector; The embedded vector is transformed into an ordinal mode to obtain a symbol sequence, and the initial permutation entropy is determined according to the symbol sequence.

[0010] In some possible implementations, the initial permutation entropy is weighted by the median absolute deviation to obtain the weighted permutation entropy, including: Determine the relative weight of ranking based on the median absolute deviation of the embedding vector; The initial permutation entropy is weighted according to the relative weight of the sorting to obtain the weighted permutation entropy.

[0011] In some possible implementations, reconstructing the weighted permutation entropy by performing multi-scale ordinal information reconstruction on the weighted permutation entropy to obtain the multi-scale permutation entropy includes: The similarity ranking matrix is coarse-grained based on non-overlapping windows and displacement offsets to obtain hierarchical ordinal information; The weighted permutation entropy is reconstructed according to the hierarchical ordinal information to obtain the multi-scale permutation entropy.

[0012] In some possible implementations, performing time series enhancement reconstruction on the multi-scale permutation entropy to obtain the time series multi-scale permutation entropy includes: The multi-scale permutation entropy is temporally enhanced and reconstructed based on the preset delay time control parameters to obtain the temporal multi-scale permutation entropy.

[0013] In a second aspect, the present invention provides a network application method for predicting fetal brain age, comprising: Obtaining a magnetic resonance image of the fetus's brain; Inputting the fetal brain magnetic resonance image to be tested into the fetal brain age prediction network to obtain the brain age prediction output result; The fetal brain age prediction network is determined according to any one of the above-mentioned fetal brain age prediction network training methods.

[0014] In a third aspect, the present invention provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any one of the above-mentioned fetal brain age prediction network training methods and / or the above-mentioned fetal brain age prediction network application methods.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the fetal brain age prediction network training method and / or the fetal brain age prediction network application method according to any one of the above items are implemented.

[0016] The beneficial effect of adopting the above embodiment is that in the fetal brain age prediction network training method provided by the present invention, the loss function constructed by the permutation entropy regularizer can capture the continuous ordered relationship in the regression, significantly improve the discrimination and generalization capabilities of the fetal brain age prediction network, and obtain higher prediction accuracy and lower mean absolute error.

[0017] Furthermore, the present invention also captures the information of changes in ordinal patterns in the embedded vector by weighting the permutation entropy by median absolute deviation, reconstructs the multi-scale ordinal information of the permutation entropy to identify the potential multi-scale ordinal patterns and time correlations between adjacent regression samples, and models the long-term and short-term structural correlations between adjacent features by performing time-series enhanced reconstruction of the permutation entropy, thereby capturing more effective supervision information in the regression and further improving the prediction accuracy of the fetal brain age prediction network. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 A schematic flow chart of an embodiment of the fetal brain age prediction network training method provided by the present invention; Figure 2 A schematic diagram of a similarity ranking conversion process according to an embodiment of the present invention; Figure 3 Schematic diagram of the process of calculating initial permutation entropy according to an embodiment of the present invention; Figure 4 Schematic diagram of the process of weighting the median absolute deviation according to an embodiment of the present invention; Figure 5 Schematic diagram of the process of reconstructing multi-scale ordinal information according to an embodiment of the present invention; Figure 6 A schematic flow chart of an embodiment of the network application method for predicting fetal brain age provided by the present invention; Figure 7 This is a schematic structural diagram of an embodiment of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0020] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0021] It should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the present invention illustrate operations implemented according to some embodiments of the present invention. It should be understood that the operations of the flowcharts can be implemented out of sequence, and steps that have no logical contextual relationship can be reversed in order or implemented simultaneously. In addition, those skilled in the art, guided by the content of the present invention, can add one or more other operations to the flowcharts or remove one or more operations from the flowcharts. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different networks and / or processor systems and / or microcontroller systems.

[0022] The terms "first" and "second" in the embodiments of the present invention are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature designated as "first" or "second" may explicitly or implicitly include at least one such feature.

[0023] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0024] The present invention provides a fetal brain age prediction network training method, application method and electronic device, which are respectively described below.

[0025] Figure 1 This is a flow chart of an embodiment of the fetal brain age prediction network training method provided by the present invention, as shown in FIG. Figure 1 As shown, the fetal brain age prediction network training method includes: S101, performing similarity sorting conversion on the fetal brain magnetic resonance image and the corresponding brain age label to obtain a similarity sorting matrix, performing vector embedding reconstruction on the similarity sorting matrix to obtain an embedding vector, and determining an initial permutation entropy of the embedding vector; Among them, in order to capture the continuous and ordered relationship in regression during the network training process, the embodiment adopts permutation entropy to construct a loss function. Permutation entropy is a nonlinear dynamic method for measuring the complexity and randomness of time series. It can analyze the arrangement pattern of adjacent points in the time series to quantify the complexity of the sequence. By constructing a loss function to analyze the difference between the permutation entropy of adjacent samples in the label space and the permutation entropy of adjacent samples in the feature space, the continuous and ordered relationship in regression can be effectively captured.

[0026] In the process of calculating the permutation entropy, the embodiment first needs to perform similarity sorting conversion on the fetal brain magnetic resonance image and the brain age label respectively to obtain a similarity matrix. For the brain age label, the similarity can be calculated by the negative absolute distance, and for the fetal brain magnetic resonance image, the similarity can be calculated by the cosine similarity. Then the calculated similarity matrix is similarity sorted to obtain a similarity sorting matrix. For the similarity matrix constructed by the brain age label, it can be sorted in ascending order and a sorting conversion rule can be established. For the similarity matrix constructed by the fetal brain magnetic resonance image, it is sorted according to the established sorting conversion rule. Finally, vector space embedding reconstruction is performed, and the similarity matrix is represented in the form of a vector in the corresponding feature space and label space respectively, and the initial permutation entropy of each is calculated based on the reconstructed embedding vector.

[0027] S102, performing median absolute deviation weighting on the initial permutation entropy to obtain weighted permutation entropy, performing multi-scale ordinal information reconstruction on the weighted permutation entropy to obtain multi-scale permutation entropy, performing time series enhancement reconstruction on the multi-scale permutation entropy to obtain time series multi-scale permutation entropy, and constructing a permutation entropy regularization loss based on the time series multi-scale permutation entropy; In order to make the permutation entropy more effective in capturing feature information, the embodiment further weights the permutation entropy by first setting weights by the median absolute deviation of each embedded vector, so as to capture the variation difference information between ordinal patterns; Then, the embodiment performs coarse-graining processing on the similarity ranking matrix to construct multi-scale ordinal information, and reconstructs the weighted permutation entropy based on the multi-scale ordinal information to identify potential multi-scale ordinal patterns and temporal correlations between adjacent regression samples; Finally, the embodiment performs temporal enhancement and reconstruction on the multi-scale permutation entropy by presetting delay time control parameters to model the long-term and short-term structural correlations between adjacent features, thereby obtaining the temporal multi-scale permutation entropy. The embodiment constructs a permutation entropy regularization loss based on this multi-enhanced temporal multi-scale permutation entropy, which can capture more effective supervisory information for regression during network training.

[0028] S103. Iteratively train the initial fetal brain age prediction network according to the permutation entropy regularization loss to obtain a fetal brain age prediction network.

[0029] Among them, in the embodiment, the initial fetal brain age prediction network uses HS-ResNet50, and the Adam optimizer is used to update the network parameters during the training process. After multiple iterative training, the optimal network weight parameters of the target prediction network are obtained, and the optimized network weight parameters are loaded into the test set. The final mean absolute error and goodness of fit of the fetal brain magnetic resonance image regression prediction are calculated to verify the prediction performance of the network.

[0030] Compared with the existing technology, the fetal brain age prediction network training method provided by the present invention can capture the continuous ordered relationship in regression by constructing a loss function through a permutation entropy regularizer, significantly improving the discrimination and generalization capabilities of the fetal brain age prediction network, and obtaining higher prediction accuracy and lower mean absolute error.

[0031] Furthermore, the present invention also captures the information of changes in ordinal patterns in the embedded vector by weighting the permutation entropy by median absolute deviation, reconstructs the multi-scale ordinal information of the permutation entropy to identify the potential multi-scale ordinal patterns and time correlations between adjacent regression samples, and models the long-term and short-term structural correlations between adjacent features by performing time-series enhanced reconstruction of the permutation entropy, thereby capturing more effective supervision information in the regression and further improving the prediction accuracy of the fetal brain age prediction network.

[0032] In some embodiments of the present invention, the fetal brain magnetic resonance image is obtained by preprocessing the initial magnetic resonance image, and the preprocessing of the initial magnetic resonance image includes: The initial magnetic resonance image is subjected to key region cutting, histogram equalization, image transformation and zero filling operations in sequence.

[0033] To improve training effectiveness, the initial MRI images are first segmented and saved for important regions. Histogram equalization is then used to ensure a uniform grayscale probability density distribution for each fetal brain MRI image, resulting in clearer images. To improve numerical quality, the embodiment also performs image transformation and zero padding. Image transformations include image rotation (0 to 360°), height and width translation (0.2), scaling (96×96), horizontal and vertical flipping, and zero padding. Zero padding is performed based on the longest side of each image to ensure equal height and width.

[0034] In some embodiments of the present invention, the similarity ranking matrix includes a label ranking matrix and a feature calibration matrix. Figure 2 FIG. 1 is a flow chart of similarity sorting conversion according to an embodiment of the present invention, as shown in FIG. Figure 2 As shown in the figure, the fetal brain magnetic resonance images and the corresponding brain age labels are transformed by similarity sorting to obtain a similarity sorting matrix, including: S201, constructing a similarity matrix for the fetal brain magnetic resonance image and the corresponding brain age label to obtain a feature similarity matrix and a label similarity matrix; S202, arranging the label similarity matrix in ascending order to obtain a label sorting matrix, and determining a sorting conversion rule based on the label sorting matrix; S203 , calibrating the feature similarity matrix according to the sorting conversion rule to obtain a feature calibration matrix.

[0035] Specifically, in order to ensure that the calculated permutation entropy can effectively capture the continuous ordered relationship in the regression, it is necessary to perform a similarity sorting transformation operation on the input brain magnetic resonance images and brain age labels.

[0036] In the embodiment, represents a subset of each batch during training, in the label space, By applying a similarity function between two age labels (e.g. negative absolute distance) to calculate the pairwise similarity matrix. Therefore, Middle Item elements can be represented as:

[0037] Then, yes Sort rows in ascending order and Get the corresponding similarity sorting transformation rule for each row in , which is defined as follows:

[0038] in, Operation to ensure sorting The similarity elements of each row in are in an ordered state, so that the permutation entropy is zero. Similarly, in the feature space, represents the feature extractor, represents the input fetal brain image, Indicates the extracted fetal brain sulcus features, By applying the similarity function (e.g. cosine similarity) to compute the pairwise similarity matrix. Middle Item elements can be defined as:

[0039] In order to maintain The consistency of similarity ranking transformation in the calculation is obtained Required application The similarity sorting conversion rules are calibrated, and the calibrated It can be expressed as:

[0040] In some embodiments of the present invention, Figure 3 Schematic diagram of the process of calculating the initial permutation entropy according to an embodiment of the present invention, as shown in FIG. Figure 3 As shown, the similarity ranking matrix is reconstructed by vector embedding to obtain the embedded vector, and the initial permutation entropy of the embedded vector is determined, including: S301, reconstructing the delayed embedding state of the similarity sorting matrix to obtain an embedding vector; S302: Perform ordinal mode conversion on the embedded vector to obtain a symbol sequence, and determine an initial permutation entropy based on the symbol sequence.

[0041] Specifically, in the vector space embedding reconstruction, the similarity ranking matrix to be reconstructed is defined as , the embodiment obtains the corresponding embedding vector by performing a delayed embedding state reconstruction operation on , and the reconstruction process can be defined as:

[0042] in, Indicates the delay time, Represents the reconstructed dimensional delay vectors, these reconstructed vectors overlap with each other, dividing the input data into Partitions. Among them, The higher the value of , the more information about the past is contained in subsequent vectors.

[0043] After completing the vector space embedding reconstruction, for each reconstructed Vector of dimensions , can be achieved through The real values are ranked and converted into an ordinal pattern, and each ordinal pattern is mapped to a specific symbol. It should be noted that Is the length of of One of the possible permutations of symbols, where , and each Each corresponds to a unique . Then, the relative frequency of each symbol can be calculated:

[0044] in, Indicates when The ordinal pattern is Type, count once, No more than , and if and only if there is no sequence of symbols of the given type When the vector Then, the ordinal-aware permutation entropy (PE) is defined as the Shannon entropy of the probability distribution of all possible permutation types extracted during the symbolization process:

[0045] When the probability of all symbol permutations is equal, the permutation entropy will reach its maximum value. On the contrary, when there is only one It is not zero, indicating that the data sequence is completely ordered and the value of permutation entropy reaches the minimum. Therefore, the permutation entropy can be further normalized to ,in , and the smaller the value of permutation entropy is, the more ordered the data sequence will be.

[0046] At this point, according to the permutation entropy, the initial loss function of the basic version, that is, the deep structured permutation entropy regularizer, can be expressed as:

[0047] in, Indicates the first OK, express norm, which is used to penalize the difference between the permutation entropy of adjacent samples in the label space and the permutation entropy of adjacent samples in the feature space. In addition, due to the permutation The similarity elements of each row in are ordered, so The permutation entropy of is zero.

[0048] In some embodiments of the present invention, Figure 4 Schematic diagram of the process of weighting the median absolute deviation in an embodiment of the present invention, such as Figure 4 As shown, the initial permutation entropy is weighted by the median absolute deviation to obtain the weighted permutation entropy, including: S401, determining the relative weight of the ranking according to the median absolute deviation of the embedding vector; S402: weighting the initial permutation entropy according to the relative ranking weight to obtain a weighted permutation entropy.

[0049] Specifically, the weighting process is performed in the step of calculating the relative frequency of each symbol, which can be expressed as:

[0050]

[0051] in, No more than , and if and only if there is no permutation of the given type Vector hour, . is from the corresponding vector The median absolute deviation of ), the median absolute deviation can capture the variation difference information between ordinal patterns. Often referred to as standardized . The weighted permutation entropy is then defined as follows:

[0052] Among them, the value of weighted permutation entropy is also located at In the interval, based on weighted permutation entropy Being able to distinguish embedding vectors with the same symbol arrangement but different variations can capture richer age-related ordinal information, thereby helping the model learn more accurate continuous and ordered representations.

[0053] In some embodiments of the present invention, Figure 5 FIG. 1 is a flow chart of multi-scale ordinal information reconstruction according to an embodiment of the present invention. Figure 5 As shown in Figure 2, the weighted permutation entropy is reconstructed by multi-scale ordinal information to obtain the multi-scale permutation entropy, including: S501, performing coarse-graining processing on the similarity sorting matrix based on non-overlapping windows and displacement offsets to obtain hierarchical ordinal information; S502 : Reconstruct the weighted permutation entropy according to the hierarchical ordinal information to obtain the multi-scale permutation entropy.

[0054] Specifically, for the similarity element of the input , the embodiment adopts a key coarse-graining technique that aggregates the The similarity information of adjacent elements in . The coarse-graining process is defined as follows:

[0055] in, represents the scale factor, Indicates not greater than The largest integer after The obtained elements can be expressed as The coarse-graining process The potential multi-scale ordinal patterns and temporal correlations between adjacent regression samples can be identified. Then, in order to further extract richer multi-scale information and enhance the exchange of ordinal information between adjacent data, a displacement offset is introduced in the coarse-graining process. , the coarse-graining process after refinement is defined as follows:

[0056] For a given scale factor ,exist The hierarchically decomposed elements embed multi-scale ordinal information. The obtained elements can be re-expressed as Next, after the coarse-graining operation, each scale The corresponding weighted permutation entropy can be defined as follows:

[0057] in, , It is a symbol The weighted relative frequency of , By calculating different levels Down-weighted relative frequency The average value of A sequence of hierarchically aggregated elements The powerful ability of ordinal information in the model improves the ordinal consistency between regression samples learned at the same scale.

[0058] Then, according to the weighted reconstructed permutation entropy, the reconstruction loss function is the deep structured permutation entropy regularizer, which is expressed as:

[0059] in, Control and balance The middle level ordinal information and the original element sequence The multi-scale permutation entropy (MPE) has a strong ability to capture multi-scale ordinal patterns. Through the enhancement of MPE, the loss function can identify the multi-scale ordinal information in the fetal brain and preserve the corresponding ordinal properties in the learned feature space. Finally, the loss function based on MPE enhancement can be defined as:

[0060] In some embodiments of the present invention, performing time series enhancement reconstruction on the multi-scale permutation entropy to obtain the time series multi-scale permutation entropy includes: The multi-scale permutation entropy is temporally enhanced and reconstructed based on the preset delay time control parameters to obtain the temporal multi-scale permutation entropy.

[0061] Specifically, to ensure the ordered consistency between the learned representation and the target, the embodiment performs delayed embedding state reconstruction operations at different delay times to encode the long-term and short-term structural correlations of the fetal brain in parallel. The formula is as follows:

[0062] in, , Controls the distance change of the delay time, Represents the scale factor. The time-series enhanced multi-scale permutation entropy TMPE models the long-term and short-term structural correlations between adjacent features, making the perception of fetal brain development characteristics more sensitive at the multi-scale time-series level. The final loss function is defined as:

[0063] When the difference between the temporal multi-scale permutation entropy TMPE calculated by the input similarity variable in the label space and the feature space approaches zero, The convergence to 0 indicates that the ordinal pattern of adjacent samples in the feature space completely matches the ordinal pattern of adjacent samples in the label space, and also indicates that the order of samples in the feature space and the order of samples in the label space are consistent. In this way, the embodiment provides an ordinal pattern related to brain age for the fetal brain age estimation regression task.

[0064] In addition, to verify the effectiveness of this solution, ten groups of random experiments were set up in the embodiment to test and compare the method of this solution with other methods, and the comparison results shown in Table 1 and Table 2 were obtained.

[0065] Table 1: Performance test comparison results of baseline models

[0066] Table 2: Test comparison results of introducing different loss functions

[0067] Among them, MAE represents the mean absolute error, R 2 represents the goodness of fit, RMSE represents the root mean square error, MAPE represents the mean absolute percentage error, and GM represents the geometric mean. It can be seen that before the introduction of the improved permutation entropy regularization loss function, the baseline model HS-ResNet did not have a significant advantage over other models in terms of test results. However, after introducing different loss functions, the permutation entropy regularization loss of this scheme significantly outperformed existing technologies in all evaluation metrics compared to other methods, verifying the effectiveness of this scheme.

[0068] In summary, in the fetal brain age prediction network training method provided by the present invention, the loss function constructed by the permutation entropy regularizer can capture the continuous ordered relationship in the regression, significantly improve the discrimination and generalization capabilities of the fetal brain age prediction network, and obtain higher prediction accuracy and lower mean absolute error.

[0069] Furthermore, the present invention also captures the information of changes in ordinal patterns in the embedded vector by weighting the permutation entropy by median absolute deviation, reconstructs the multi-scale ordinal information of the permutation entropy to identify the potential multi-scale ordinal patterns and time correlations between adjacent regression samples, and models the long-term and short-term structural correlations between adjacent features by performing time-series enhanced reconstruction of the permutation entropy, thereby capturing more effective supervision information in the regression and further improving the prediction accuracy of the fetal brain age prediction network.

[0070] Furthermore, the present invention also provides a fetal brain age prediction network application method, combined with Figure 6 Come and see, Figure 6 This is a flow chart of an embodiment of the network application method for predicting fetal brain age provided by the present invention, as shown in FIG. Figure 6 As shown, the fetal brain age prediction network application method includes: S601, obtaining a magnetic resonance imaging image of the fetus' brain; S602, inputting the magnetic resonance image of the fetal brain to be tested into the fetal brain age prediction network to obtain a brain age prediction output result; The fetal brain age prediction network is determined according to the above-mentioned fetal brain age prediction network training method.

[0071] In an embodiment of the present invention, first, a brain magnetic resonance image of the fetus to be tested is effectively acquired, and then the fetal brain age is effectively predicted using the above-mentioned fully trained fetal brain age prediction network to output a brain age prediction output result.

[0072] like Figure 7As shown, the present invention also provides an electronic device 700 , which includes a processor 701 , a memory 702 and a display 703 . Figure 7 Only some of the components of the electronic device 700 are shown, but it should be understood that implementing all of the shown components is not a requirement, and more or fewer components may alternatively be implemented.

[0073] In some embodiments, the memory 702 may be an internal storage unit of the electronic device 700, such as a hard disk or memory of the electronic device 700. In other embodiments, the memory 702 may also be an external storage device of the electronic device 700, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 700.

[0074] In some embodiments, the processor 701 can be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code or process data stored in the memory 702, such as the fetal brain age prediction network training method and / or fetal brain age prediction network application method in the present invention.

[0075] In some embodiments, display 703 can be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 703 is used to display information about electronic device 700 and to display a visual user interface. Components 701-703 of electronic device 700 communicate with each other via a system bus.

[0076] In one embodiment, when the processor 701 executes the fetal brain age prediction program in the memory 702, the following steps may be implemented: The fetal brain magnetic resonance image is input into the initial fetal brain age prediction network to obtain the brain age prediction distribution, and the prediction loss is determined based on the brain age prediction distribution and the corresponding actual brain age; Perform similarity ranking transformation on the fetal brain magnetic resonance images and the corresponding brain age labels to obtain a similarity ranking matrix, perform vector embedding reconstruction on the similarity ranking matrix to obtain an embedding vector, and determine the initial permutation entropy of the embedding vector; The initial permutation entropy is weighted by the median absolute deviation to obtain the weighted permutation entropy, the weighted permutation entropy is reconstructed by multi-scale ordinal information to obtain the multi-scale permutation entropy, the multi-scale permutation entropy is reconstructed by time series enhancement to obtain the time series multi-scale permutation entropy, and the permutation entropy regularization loss is constructed based on the time series multi-scale permutation entropy; The initial fetal brain age prediction network was iteratively trained according to the permutation entropy regularization loss to obtain the fetal brain age prediction network.

[0077] And / or implement the following steps: Obtaining a magnetic resonance image of the fetus's brain; The magnetic resonance imaging (MRI) image of the fetal brain to be tested is input into the fetal brain age prediction network to obtain the brain age prediction output result.

[0078] It should be understood that, when the processor 701 executes the fetal brain age prediction program in the memory 702 , in addition to the above functions, it can also implement other functions. For details, please refer to the description of the corresponding method embodiment above.

[0079] Accordingly, an embodiment of the present application also provides a computer-readable storage medium, which is used to store computer-readable programs or instructions. When the program or instructions are executed by a processor, it can implement the steps or functions of the fetal brain age prediction network training method and / or fetal brain age prediction network application method provided in the above-mentioned method embodiments.

[0080] Those skilled in the art will appreciate that all or part of the process flow of the above-described method embodiment can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program, and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0081] The fetal brain age prediction network training method, application method and electronic device provided by the present invention are introduced in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A fetal brain age prediction network training method, characterized in that: include: performing similarity ranking conversion on the fetal brain magnetic resonance image and the corresponding brain age label to obtain a similarity ranking matrix, performing vector embedding reconstruction on the similarity ranking matrix to obtain an embedding vector, and determining an initial permutation entropy of the embedding vector; Performing median absolute deviation weighting on the initial permutation entropy to obtain weighted permutation entropy, performing multi-scale ordinal information reconstruction on the weighted permutation entropy to obtain multi-scale permutation entropy, performing time series enhancement reconstruction on the multi-scale permutation entropy to obtain time series multi-scale permutation entropy, and constructing a permutation entropy regularization loss based on the time series multi-scale permutation entropy; The initial fetal brain age prediction network is iteratively trained according to the permutation entropy regularization loss to obtain a fetal brain age prediction network.

2. The fetal brain age prediction network training method according to claim 1, characterized in that: The fetal brain magnetic resonance image is obtained by preprocessing the initial magnetic resonance image, wherein the preprocessing of the initial magnetic resonance image includes: The initial magnetic resonance image is subjected to key region cutting, histogram equalization, image transformation and zero filling operations in sequence.

3. The fetal brain age prediction network training method according to claim 1, characterized in that: The similarity ranking matrix includes a label ranking matrix and a feature calibration matrix. The similarity ranking matrix is obtained by performing similarity ranking conversion on the fetal brain magnetic resonance image and the corresponding brain age label, including: A similarity matrix is constructed for the fetal brain magnetic resonance images and the corresponding brain age labels to obtain a feature similarity matrix and a label similarity matrix; Arrange the label similarity matrix in ascending order to obtain a label sorting matrix, and determine a sorting conversion rule according to the label sorting matrix; The feature similarity matrix is calibrated according to the sorting conversion rule to obtain a feature calibration matrix.

4. The fetal brain age prediction network training method according to claim 1, characterized in that: The performing vector embedding reconstruction on the similarity ranking matrix to obtain an embedded vector and determining an initial permutation entropy of the embedded vector includes: Performing delayed embedding state reconstruction on the similarity ranking matrix to obtain an embedding vector; An ordinal mode conversion is performed on the embedded vector to obtain a symbol sequence, and an initial permutation entropy is determined according to the symbol sequence.

5. The fetal brain age prediction network training method according to claim 1, characterized in that: The step of weighting the initial permutation entropy by median absolute deviation to obtain weighted permutation entropy includes: Determining a relative ranking weight based on the median absolute deviation of the embedding vector; The initial permutation entropy is weighted according to the relative ranking weight to obtain a weighted permutation entropy.

6. The fetal brain age prediction network training method according to claim 1, characterized in that: The step of reconstructing the weighted permutation entropy by performing multi-scale ordinal information reconstruction on the weighted permutation entropy to obtain the multi-scale permutation entropy includes: Performing coarse-graining processing on the similarity ranking matrix based on non-overlapping windows and displacement offsets to obtain hierarchical ordinal information; The weighted permutation entropy is reconstructed according to the hierarchical ordinal information to obtain the multi-scale permutation entropy.

7. The fetal brain age prediction network training method according to claim 1, characterized in that: The performing time series enhancement and reconstruction on the multi-scale permutation entropy to obtain time series multi-scale permutation entropy includes: The multi-scale permutation entropy is subjected to time series enhancement and reconstruction based on a preset delay time control parameter to obtain a time series multi-scale permutation entropy.

8. A network application method for predicting fetal brain age, characterized in that: include: Obtaining a magnetic resonance image of the fetus's brain; Inputting the magnetic resonance image of the fetal brain to be tested into a fetal brain age prediction network to obtain a brain age prediction output result; Wherein, the fetal brain age prediction network is determined according to the fetal brain age prediction network training method according to any one of claims 1 to 7.

9. An electronic device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the fetal brain age prediction network training method according to any one of claims 1 to 7 and / or the fetal brain age prediction network application method according to claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the fetal brain age prediction network training method according to any one of claims 1 to 7 and / or the fetal brain age prediction network application method according to claim 8 are implemented.

Images